satuke/s1-mini-coreai-ios

S1-mini by Superwhisper — Core AI bundle for iOS 27 (Neural Engine, 8-bit)

This is "S1-mini" by "Superwhisper" (superwhisper/s1-mini), a Qwen3-0.6B fine-tune for cleaning up speech-to-text transcripts, converted to an Apple Core AI bundle so it runs on the iPhone Neural Engine (iOS 27+). Used experimentally by the noboard voice keyboard.

Changes from the original (Apache-2.0 §4(b))

• Exported with Apple's coreai-models recipe: coreai.llm.export superwhisper/s1-mini --platform iOS --experimental --compute-precision float16 with 8-bit k-means palettization (per_grouped_channel, group size 8; embeddings not palettized).
• tokenizer/chat_template.jinja: thinking is off unless enable_thinking is explicitly true (the model must never think).
• Weights are otherwise the original model's, quantized.

Apple's mixed 4/8-bit recipe badly degraded this model; this 8-bit export matched the original on our 66-case dictation bench (4.20 vs 4.18 / 5).

Use

Load with CoreAILanguageModel(resourcesAt:) from coreai-models and use S1-mini's own system prompt and control line from its model card.

License

Apache License 2.0 plus Superwhisper's additional term (see LICENSE): any use or distribution must keep identifying the model as "S1-mini" by "Superwhisper". Base model Qwen3-0.6B, Copyright 2024 Alibaba Cloud, Apache-2.0.